Paraísos Fiscales, Wealth Taxation, and Mobility...2021/01/20 · acknowledges ﬁnancial support...

Paraísos Fiscales, Wealth Taxation, and Mobility∗

David R. Agrawal† Dirk Foremny‡ Clara Martínez-Toledano§

December 2020

Abstract

This paper analyzes the effect of wealth taxation on mobility and the consequences fortax revenue and wealth inequality. We exploit the unique decentralization of the Spanishwealth tax system in 2011—after which all regions levied positive tax rates except forMadrid—using linked administrative wealth and income tax records. We find that fiveyears after the reform, the stock of wealthy individuals in the region of Madrid increasesby 10% relative to other regions, while smaller tax differentials between other regions donot matter for mobility. We rationalize our findings with a theoretical model of evasionand migration, which suggests that evasion is the mechanism most consistent with allof the mobility response being driven by the paraíso fiscal. Combining new subnationalwealth inequality series with our estimated elasticities, we show that Madrid’s statusas a tax haven reduces the effectiveness of raising tax revenue and exacerbates regionalwealth inequalities.

Keywords: Wealth Taxes, Mobility, Inequality, Enforcement, Fiscal Decentralization,Tax Havens, Evasion

JEL: E21, H24, H31, H73, J61, R23

∗The paper benefited from comments by Arun Advani, Facundo Alvaredo, Olympia Bover, Marius Brül-hart, Marta Espasa, Gabrielle Fack, Trevor Gallen, Nathaniel Hendren, James Hines, William Hoyt, Hen-rik Kleven, Camille Landais, Juliana Londoño-Vélez, Mohammed Mardan, Mariona Mas, Mathilde Muñoz,Thomas Piketty, Daniel Reck, Emmanuel Saez, Guttorm Schjelderup, Kurt Schmidheiny, Joel Slemrod,Daniel Waldenström, David Widlasin, Owen Zidar, James Ziliak, Floris Zoutman, Gabriel Zucman, EricZwick, as well as seminar participants at the Center for European Economic Research (ZEW), InternationalInstitute of Public Finance, International Online Public Finance Seminar, Norwegian School of Economics,National Tax Association, Paris School of Economics, University of Girona, University of Barcelona, Uni-versidad de La República Uruguay (iecon) and University of California, Los Angeles. Foremny gratefullyacknowledges financial support from Fundación Ramón Areces and grant RTI2018-095983-B-I00 from theMinisterio de Ciencia, Innovación y Universidades; and Martínez-Toledano from Fundación Rafael del Pino.Any remaining errors are our own.†University of Kentucky, Department of Economics and Martin School of Public Policy & Administration,

433 Patterson Office Tower, Lexington, KY 40506-0027; email: [email protected]; phone: 001-859-257-8608.Agrawal is also a Fellow of CESifo.‡Department of Economics and I.E.B., Universitat de Barcelona, Facultat d’Economia i Empresa - Av.

Diagonal, 690 (08034 Barcelona) Spain; e-mail: [email protected]; phone +34 93 402 18 16. Foremny is alsoan affiliate member of CESifo.

§Columbia Business School, Uris Hall, 3022 Broadway, New York, NY, 10027, United States; email:[email protected]; phone: 001-646-520-6307.

Electronic copy available at: https://ssrn.com/abstract=3676031

Rising capital shares of income and the associated increases in inequality observed inmany countries have spurred new interest in the taxation of wealth. Many policy discussionshave focused on whether wealth taxes are enforceable, as taxpayers might respond to wealthtaxes by moving assets to tax havens.1 Indeed, empirical evidence finds that a significantfraction of financial assets owned by the wealthy is held offshore (Alstadsæter et al., 2019).The risk of tax-induced mobility was a motivating factor in Piketty (2014)’s call for a globalwealth tax: “if not all countries implement a wealth tax, then mobile capital would simplyflow to tax havens where wealth tax rates are zero.”

Analyzing the mobility responses to wealth taxes is, however, an empirical challenge.Wealth taxes provide limited sources of exogenous variation, as they are often implementedat the national level and for the very top of the wealth distribution. Given the difficultyof cross-country comparisons, little variation in wealth taxes exists across individuals orregions within a country, and when they do exist, they often do not feature a prominent taxhaven. Furthermore, any study of migration must know where the taxpayer originated fromand migrated to, which requires potential harmonization of multiple countries’ or regions’administrative tax records. Thus, despite the importance of an annual wealth tax in recentpolicy and academic debates, important questions necessary to evaluate its suitability remainunanswered. How large are the mobility responses to wealth taxation and what role do taxhavens play? Are these mainly avoidance (i.e., real migration) or evasion (i.e., fraudulentdeclaration of fiscal residence) responses? How do these responses shape wealth tax revenuesand wealth inequality dynamics?

We break new ground on these issues by using arguably exogenous variation in wealthtax rates across sub-national regions (Comunidades Autónomas) within Spain. Prior to 2008,Spain had a mostly uniform wealth tax, which was briefly suppressed. It is only after itsreintroduction in 2011 that regions started to substantially exercise their autonomy to changewealth tax schedules. As a consequence, large differences in effective tax rates emerged acrossregions under this residence-based tax system. Madrid plays a special role in this setting asan internal paraíso fiscal with a zero effective tax rate on wealth. The presence of this salienttax haven distinguishes Spain from one other country with decentralized wealth taxes—Switzerland—where the variation results from tax rate differentials, but with all regions

1Recent political debates, including in the United States, have centered around the wealth tax as a revenuesource to fund public programs and to reduce wealth inequality. Beyond national proposals, states such asCalifornia have proposed decentralized state-level wealth taxes (Gamage et al., 2020).

1


levying positive tax rates due to federal restrictions that prevent regions to abolish the tax.2

This distinction is critical to testing Piketty’s claim that tax havens play a special role inundermining wealth taxation. Moreover, the presence of a single zero-tax region internal toSpain—-likely to arise in decentralized federations considering wealth taxes in the absence offederal restrictions—creates unique incentives for escaping wealth taxation that do not arisein a setting with small tax differentials, but no salient tax haven.

To conduct the analysis, we assemble administrative wealth tax records for a region-ally stratified and longitudinal random sample prior to the suppression of the wealth tax(2005-2007) and merge them to administrative personal income tax records before and afterdecentralization (2005-2015). The individual personal income tax records contain informa-tion on fiscal residence, which is unique to all personal taxes, making it possible to follow thelocation of wealth tax filers before and after decentralization.

[Figure 1 about here.]

The key result of our paper can be seen in the raw administrative data (Figure 1). Weplot the change in the number of individuals who would be subject to the post-reform wealthtax for Madrid and the average of the other regions. Following Madrid’s decision to become atax haven, the number of wealth tax filers reporting Madrid as their fiscal residence increasesby over 6,000. The other regions see an average decline of 375 filers. Relative to 2010, thisrepresents an approximately 10% increase in the stock of wealth tax filers in Madrid.

To analyze the effect of wealth taxation on mobility, we proceed in three steps. First, wepresent descriptive evidence on the number of movers between all pairs of Spanish regions.Following decentralization, the number of wealth tax filers moving to Madrid is substantiallyhigher than the number of moves to any other region, including larger regions.

Second, we aggregate the individual data to the region-year-wealth tax filer level andcompare the population of wealth tax filers in Madrid to the population of wealth tax filersin other regions. We find a 10% increase in the relative population in Madrid by five yearsafter decentralization of the wealth tax. Identification follows from a difference-in-differencesdesign. Any threat to identification would come from a shock that makes Madrid relativelymore attractive compared to other regions. To address this, we add an additional differenceexploiting information on the relative population of wealth tax filers and high capital incomeindividuals not subject to the wealth tax. Thus, any shock threatening our results must only

2Hines (2010) and Hines and Rice (1994) define tax havens as jurisdictions that have low tax rates orloopholes on particular assets and self-promote themselves as a center for those assets, which Madrid satisfies.Furthermore, the popular press and politicians have dubbed Madrid a tax haven. Madrid, however, is unlikemuch of the stereotypical tax competition and tax havens literature (Hines, 2010; Dharmapala and Hines,2009; Kessler and Hansen, 2001), where low-tax jurisdictions are small.

2


affect wealth tax filers, but not high capital income non-filers. We document that non-filers donot view Madrid any more attractive after the reform. Furthermore, we show that migrationeffects follow tax changes and do not predate them; there are no significant pretrends inthe periods prior to the reform. Furthermore, most wealth tax filers are either pure rentiersor have limited labor income, so that Spain’s regional differences on labor income taxes areirrelevant. The elasticity with respect to the net-of-tax rate on wealth is 7.9, which translatesto an elasticity with respect to a capital income tax of 0.34. This elasticity is in the rangeof short-run elasticities in the income-tax literature (Kleven et al., 2020).

Finally, in addition to the aggregate analysis, we exploit an orthogonal source of variationrelying on the progressivity of the wealth tax in the context of an individual location choicemodel. This specification allows us to account for region-by-year fixed effects, which controlfor shocks that may influence preferences for a particular region in a particular year, suchas time varying amenities or other regional policies. This approach also accounts for fixedcharacteristics of the mover that are constant across alternative regions, for any sorting basedon characteristics, and for other policy changes that affect all wealth tax filers. Further, thismodel allows us to analyze heterogeneous effects across individuals. In line with the aggregateanalysis, we find that only the tax rate of Madrid matters for relocation choices: Madrid’spopulation increases due to its zero tax rate, but the populations of regions with low butpositive tax rates remain approximately unchanged. There is little heterogeneity acrossindividual characteristics. We only find larger effects at the top of the wealth distribution,as the incentives to escape wealth taxation are higher the further up one moves along thewealth distribution due to the progressivity of the tax.

To shed light on the mechanisms behind the mobility responses, we build a simple the-oretical model in which taxpayers have the choice over migrating or evading. In a standardmobility model without evasion, even a small tax differential will attract some individuals atthe margin. However, in the presence of evasion, if audit probabilities are sufficiently small,an individual who finds it advantageous to evade will never find it optimal to falsely declarea region other than the tax haven. Given our empirical analysis shows that almost all fiscalresidence changes involve Madrid, the theoretical model indicates our results are likely drivenby evasion rather than real responses. Further, we then digitize region-specific wealth taxaudit records and correlate the audit rates with mobility changes. In standard models of eva-sion, the audit probability increases with the amount evaded. Consistent with this, we showthat audit rates are positively correlated with mobility to Madrid, but not with mobility toother regions, suggesting that the tax authority believes that most fraudulent moves involveMadrid. Taken together, evasion is the dominant mechanism behind fiscal residence changes.

We then use our estimates to study the effect of tax-induced mobility on wealth and

3


income tax revenues by means of counterfactual simulations. We simulate the evolution ofwealth and income tax revenue absent tax-induced mobility, a scenario that could potentiallybe achieved by extensive enforcement, tax harmonization, or minimum tax rates. We findthat Spain foregoes on average 5% of total wealth tax revenue due to tax-induced mobility,with substantial differences across regions. We also find important differences in foregoneincome tax revenue across regions due to tax-induced mobility, but little income tax revenueis foregone at the national level.

An unresolved theoretical debate is whether tax harmonization or minimum tax ratesare Pareto improving (Kanbur and Keen, 1993). To shed light on this, we also simulatethe evolution of revenue under a centralized wealth tax system in which all regions wouldhave a uniform wealth tax schedule and a system with minimum tax rates. Abolishing thedecentralized system in favor of a centralized system leads to large revenue gains mainly dueto the added tax revenue from taxing the base in Madrid. However, we show that this is nota Pareto improvement unless harmonization is to a rate that is very close to the maximumdecentralized rate. Minimum tax rates could instead increase revenue in all regions.

Finally, we study the interplay between the observed mobility responses and regionalwealth inequality dynamics.3 To do this, we build new top national and regional wealthdistribution series. The main novelty is that we decompose the wealth shares at the regionallevel: this is the first attempt to construct harmonized top wealth shares across sub-nationalregions. Most prior studies of spatial inequality focus on income inequality, economic op-portunity or poverty—not on wealth inequality—and emphasize the importance of analyzingspatial variation along with socioeconomic variables to determine optimal policy responses(Chetty and Hendren, 2018b; Chetty and Hendren, 2018a). Our new regional wealth dis-tribution series reveal the existence of significant differences in both the level and trend inwealth concentration across Spanish regions.

We take advantage of the regional wealth series to simulate the spatial dynamics of wealthinequality absent tax-induced mobility. The mobility of wealthy taxpayers to Madrid has ledto a significant rise in wealth concentration in the region. In particular, between 2010 and2015 the top 1% wealth share growth rate in Madrid (16%) was almost double the growth ratehad tax-induced mobility not existed (8.7%). This finding contrasts with the decline in thetop 1% wealth share in the rest of Spain after decentralization. Overall, these results revealthat Madrid’s status as a tax haven has exacerbated regional wealth inequalities. Even though

3For the literature on wealth inequality and progressive wealth taxation, please see Kopczuk and Saez (2004)Piketty and Saez (2014), Piketty and Zucman (2014), Kopczuk (2013), Kopczuk (2015), Jones (2015), Saezand Zucman (2016), Smith et al. (2019b), Saez and Zucman (2019a), and Kopczuk (2019). Alvaredo andSaez (2009) document wealth inequality in Spain. See Bonhomme and Hospido (2013) and Bonhomme andHospido (2017) for measures of income inequality in Spain.

4


much of the mobility is due to tax evasion, increases in regional wealth inequality are relevantas they are highly correlated with political influence (Gilens and Page, 2014). Individualswho change their fiscal domicile to Madrid via evasion do not care about the provision ofpublic services, but instead may lobby for policies that are not necessarily aligned with thepreferences of residents by protecting their own financial interests via lower taxes.

This paper contributes to three main strands of the literature. First, the nascent empiricalliterature studying behavioral responses to wealth taxation is based on estimation of taxablewealth elasticities, placing an upper bound on the mobility elasticity (Kleven et al., 2020).4

One exception is Brülhart et al. (2016), which uses data for Swiss cantons and shows thatobserved cross-canton responses can mostly be attributed to changes in wealth holdingsrather than mobility across localities in Switzerland. However, all Swiss cantons (must) levya positive wealth tax. Our paper is thus the first to study mobility responses to a tax haven,or put differently, adopting or not adopting a wealth tax. Such analysis is critical to testingPiketty’s claim that a wealth tax cannot be successful unless globally adopted.

Second, our work also relates to the literature on the effect of taxes on mobility. Individu-als (and wealth) moving across borders may threaten the ability to engage in redistribution orto raise revenue. Given that top-taxpayers contribute a disproportionate share of taxes, muchof the literature has focused on top-income earners.5 However, the literature on wealth-taxinduced mobility is scant, and there is no evidence about how these responses might shapewealth inequalities across receiving and sending regions.6 We contribute to this literature byusing our wealth tax-induced mobility responses to study the impact of tax havens, and theresulting mobility induced therein, on the dynamics of the wealth distribution.

Finally, our results also have important implications for the literature on tax enforcementand the ability of governments to raise revenue from decentralized capital taxes. Recentempirical evidence has shown that behavioral responses of high wealth individuals dependon the enforcement environment (Slemrod, 2019, Londoño-Vélez and Ávila-Mahecha, 2020).The link between our mobility estimates and the dynamics of tax revenue sheds new lighton the importance of the degree of enforcement and the extent of fiscal decentralization. In

4With respect to the elasticity of taxable wealth, studies generally find large effects: Jakobsen et al. (2020)use administrative wealth records from Denmark; Zoutman (2016) for The Netherlands; Seim (2017) forSwedish wealth tax payers, Londoño-Vélez and Ávila-Mahecha (2020) for Colombia, and Durán-Cabré etal., 2019 for Catalonia. This literature generally does not focus on off-shoring of wealth and its mobility.

5Although estimates vary, at the margin, taxes appear to be a factor in the location choices of top earners(Agrawal and Foremny, 2019, Akcigit et al., 2016, Kleven et al., 2013, Kleven et al., 2014, Schmidheiny andSlotwinski, 2018, Moretti and Wilson, 2017, Muñoz, 2019, Young and Varner, 2011, Young et al., 2016).

6There is a nascent literature on bequest and estate taxes suggesting that the location decisions of the elderlyare not very responsive (Brülhart and Parchet, 2014, Bakija and Slemrod, 2004, Conway and Rork, 2006),except at the very top (Moretti and Wilson, 2019). Brülhart et al. (2016) decompose the elasticity of taxablewealth and find that mobility accounts for one quarter of the effect.

5


particular, our decentralized setting allows us to compare several policy reforms—tax har-monization, minimum tax rates, and increased enforcement—debated theoretically, but forwhich empirical comparisons of their effects are nonexistent. Although our estimates sug-gest that decentralized wealth taxation is possible in the short-run, consistent with Piketty’scall for a global wealth tax, the case of decentralized taxation in Spain falls victim to thepresence of a fiscal paradise. Although a centralized tax would be subject to mobility op-portunities external to the country, it could be coupled with more aggressive enforcementmechanisms. As an alternative to centralization, our results show that any coordinated taxschedule that will have the political support of all regions must be at a tax rate sufficientlyclose to the maximum rate. Absent a political consensus to harmonizes the wealth tax,appropriate enforcement measures must be in place: centrally imposed minimum tax rates,increased auditing and information sharing between the central and regional governments, orthe taxation of immobile assets such as land according to the source-principle. Nonetheless,increases in enforcement may only be met with modest success (Johannesen et al., 2020), asenforcement mechanisms may induce new evasion strategies.

1 Institutional Details

The Spanish wealth tax was introduced in 1978 (Law 50/1977) to complement the personalincome tax. The tax was briefly suppressed between 2008 and 2010. All regions are subjectto this tax except for Basque Country and Navarre, which due to their special status areautonomous to design most taxes and hence, they have their own wealth tax. The taxschedule is progressive and it is applied to the sum of all individual wealth components netof debts. Since 2011, it is only levied if net taxable wealth (i.e., taxable assets - liabilities)are above 700,000 Euro (approximately the top 0.5% of the total adult population in 2015).Over the period 2002-2007, the filing threshold was 108,182.18 Euro (approximately 2.7% ofthe total adult population in 2007). Given the tax is on individual, and not joint wealth,joint assets are split among spouses.7


Since 1997, the rights to modify the amount exempted and the tax rates were ceded to theregions, under the condition of keeping the national statutory minimum bracket and minimummarginal tax rates (default schedule). In 2002, the regions were given the right to changeor include deductions in the wealth tax and the condition of requiring a minimum bracket

7For further details, see Appendix A.1.

6


and marginal tax rates was suppressed. All regions kept the national wealth tax schedule(i.e., 0.2-2.5%) during the 1990’s and early 2000’s. In the mid-2000’s a few small changeswere implemented by some regions. Thus, it is only after its reintroduction in 2011 whensignificant differences in the wealth tax emerge. For instance, Madrid decided to keep thewealth tax suppressed after 2011, contrary to Andalusia and later regions such as Cataloniaand Extremadure who have raised the marginal tax rates above the default schedule. Thefirst panel of Figure 2 shows the marginal tax rates under the centralized wealth tax and theremaining panels show the variation in tax rates across the fifteen regions following 2011.

The reintroduction of the wealth tax was authorized in September 2011 and initially camewith substantial uncertainty over when or if it would actually be reimplemented by regionalgovernments. The authorization was sunset to only apply retroactively for 2011 and thefollowing year. Immediately after the central government’s decision, the regional governmentin Madrid announced the suppression of the wealth tax and applied a 100% tax credit. Tohave a different tax schedule than the national default, regions must actively pass a law.However, many other regions did not formulate their wealth tax schedules immediately. Thiscreated additional delays over what each region’s tax schedule would look like. In September2012, the central government announced the extension of the wealth tax until 2013 and thisprocedure continues on annually (Durán-Cabré et al., 2019).

For the purposes of this study, it is important to know the definition of fiscal residenceand to understand how taxpayers can change their fiscal residence by “moving”. The fiscalresidence is the property that constitutes the primary residence of the taxpayer and it is thesame for all personal taxes. For a property to be qualified as primary residence, the wealthtaxpayer needs to have lived there continuously over at least three years. An exception ap-plies in case of death of a family member, marriage, divorce, first job, job transfer or anyother analogous circumstance (Law 40/1998, Law 35/2006). Updating the fiscal residencefor tax purposes can be directly done on the tax form. Despite the legal regulations pre-venting the immediate change of fiscal residence, taxpayers find it easy to change their fiscalresidence either by pretending they live in a rented property, in their secondary residence(approximately 86% of wealth taxpayers had at least one secondary residence in 2010), orin the residence of a relative. Auditing falls to both the central and regional authorities.However, verifying the primary address comes with substantial administrative costs to thetax authorities. Enforcement in a multi-tier setting creates coordination problems.

The decentralization of the wealth tax should be considered in the general context offiscal decentralization in Spain. The central government also passed provisions that allowthe regions to set the tax brackets and tax rates on their half of the personal income tax onlabor. While this decentralization created incentives for high (labor) income individuals to

7


move, Spain operates a dual income tax system. Under this system, capital income is taxedat a common schedule. Thus, for high-wealth individuals who obtain a substantial fractionof their income from the return to capital, decentralization of the labor income tax providedlittle additional incentive to move. Figure A1 shows after linking administrative wealth taxto personal income tax records, approximately 80% of wealthy individuals have labor incomebelow 100,000 Euro. Critically, as shown in Agrawal and Foremny (2019), the incentives tomove due to the labor income tax are negligible for incomes below 100,000 Euro in our periodof study.8

Inheritance taxes have been decentralized to the regions since 1997, but regions did notexercise this right until the mid-2000s. In particular, Madrid adopted a tax credit of 99% onclose relatives starting already in 2007, such that there is no additional incentive created bythis tax starting in 2011. Moreover, the place of residence for this tax is defined based onthe location of the deceased over the last five years before death. Given this long durationof proof, and the fact that we focus on five years following decentralization, we expect littleof the mobility we identify to be a result of these taxes.9

2 Data

We combine two administrative data sets constructed by the Spanish Institute of FiscalStudies in collaboration with the State Agency of Fiscal Administration. We obtain theseconfidential records with approval from the Spanish government.

The first data set (Panel de Declarantes del Impuesto sobre la Renta de las PersonasFísicas, 1999-2015) is an approximately 4% sample of individual level personal income taxreturns from 1999-2015. The data contains all items reported on the annual personal incometax declaration. This includes the amount and source of income, personal characteristics(e.g., age and gender), and, critically, the region of fiscal residence of the tax filer. The panelstructure allows us to follow individuals over time. The micro-files are drawn from 15 ofthe 17 autonomous communities of Spain, in addition to the two autonomous cities, Ceutaand Melilla. Two autonomous regions, Basque Country and Navarre, are excluded, as theydo not belong to the Common Fiscal Regime. The second data set (Panel de Declarantesdel Impuesto sobre el Patrimonio, 2002-2007) includes administrative wealth tax returns,including detailed information about wealth taxpayers’ assets and liabilities. This data isavailable for individuals included in the income tax panel who were subject to the wealth tax

8To address this possible confounding event, we implement a robustness check where we eliminate all wealth-tax filers that have labor income in excess of 100,000 Euro.

9See Appendix A.1 for a more detailed discussion about capital taxes in Spain.

8


between 2002-2007. No centralized data are available after the wealth tax was suppressed.Nonetheless, as the legal definition of fiscal residence for both wealth and income taxes is thesame, we rely on the one reported in the income tax returns.

We have also been granted access to the universe of wealth tax records for Cataloniafollowing decentralization. We use this additional data for some of our robustness checks.Even if we had wealth tax information for all regions, these data would not be sufficient, asnational law only requires residents in Madrid to file a wealth tax return if their gross wealthexceeds 2,000,000 Euro.

The income tax dataset is stratified by region, income level and main source of income,and it oversamples the top of the distribution. Given this stratification, the data are meantto be representative of the personal income tax distribution. We reweight the data to berepresentative of the total population of personal income taxpayers and wealth taxpayersacross regions. First, we reweight the sample of wealth taxpayers to match regional totalsover the period 2002-2007. We then extrapolate these weights forward by applying region-specific adult-age population growth rates. Finally, we reweight the subsample of personalincome taxpayers that do not file wealth taxes so that after reweighting, the full panel matchesthe total number of personal income taxpayers in each region and year. The reweighting isimportant for the inequality analysis in Section 6.1, but does not affect our regression results.

2.1 Wealth Extrapolation Method

The main variable we use in our the analysis is the fiscal residence, which we directlyobserve in the tax records on an annual basis. However, we need to estimate wealth for theyears for which wealth tax records are not available (2008-2015) to define treatment statusand to build the wealth distribution series. We do so by combining national accounts, wealthand personal income tax returns. Following Martínez-Toledano (2020), we map each personalincome category from national accounts to a personal wealth category in non-financial andfinancial accounts.10 Then, we compute the annual rate of return for each asset category asthe ratio of the stock to the flow. Using these returns, we then extrapolate individual wealthfrom 2008 onward using reported individual wealth in 2007 as an anchor.

Asset categories for which the aggregate rate of return is not available (e.g., jewelry, an-tiques, rural real estate, industrial and intellectual property rights) are extrapolated forwardusing the annual growth rate of the average reported values from official aggregate wealth

10For non-financial accounts we rely on the reconstruction done by Artola Blanco et al. (2020) and forfinancial accounts on the Bank of Spain balance sheets. We can map urban real estate, business assets, lifeinsurance, deposits, debt assets, shares and debts.

9


tax records published by the Spanish Tax Agency over the period 2011-2015.11

To show the robustness of our extrapolation method, we first compare extrapolated av-erage regional wealth to the actual reported average wealth published by the Spanish TaxAgency. Figure A2 shows that the extrapolation matches regional average wealth in bothlevel and trend. We also compare extrapolated versus actual individual reported wealth lev-els using Catalonia’s administrative wealth tax records after 2011 (Figure A3). The strongcorrelation between our extrapolated and the direct wealth measures in this region providesevidence supporting our method. In particular, the latter figure suggests the extrapolationis extremely accurate at lower levels of wealth, but underestimates wealth at the very top.Given this extrapolation is mainly used to classify individuals near the 700,000 threshold,the underestimation at the very top is not a major concern. Indeed, we show our results arerobust to the use of the direct 2007 wealth data to define treatment status.

2.2 Tax Calculator

Research on tax-induced mobility requires knowing the tax liabilities an individual paysin their region of residence and all possible counterfactual regions of residence. As thereexists no publicly available wealth tax simulation model for Spain, we have constructed ourown tax simulator accounting for all important details of the Spanish wealth tax system.

The wealth tax simulator allows us to calculate wealth tax liabilities for every regionand year. To do this, we comb through thousands of pages of tax documents providedby the Spanish Ministry of Finance.12 In particular, we collect for each year and region,most parameters of the wealth tax code, such as rates, brackets, exemptions (e.g., primaryresidence, business assets), reductions (e.g., disability), as well as the maximum combinedwealth and income tax liability cap.13

We use the tax calculator to simulate for each individual the average tax rate in herregion of residence and hypothetical tax rates if she lived in any other region. The taxsimulator thus provides all counterfactual levels of the wealth tax burden across regions ofSpain under both a decentralized and centralized wealth tax system. For the years in whichdirect individual wealth information is available, results of the tax calculator consistentlymatch the information available in the administrative tax return data.

11For some assets (e.g., taxable business assets, liabilities), we also use this last procedure, as it better matchesthe evolution of total reported wealth by region. We refine the extrapolation by adjusting reported urbanreal estate to account for the exemption on main residence, which was raised in 2011.

12This information is released each year and summarizes all parameters with examples how to calculate thetax burden of the personal income tax and the wealth tax (Manual Práctico de Renta y Patrimonio).

13All details about the tax simulator are described in Appendix A.2.

10


2.3 Treatment and Comparison Groups

In this subsection, we define the treated and comparison individuals that we will use in thesubsequent analyses. As the treatment status must be defined using data prior to the wealthtax reintroduction, we face a trade-off of using the raw 2007 administrative data (under thecentralized regime) versus 2010 extrapolated data (under the wealth tax suppression). Forthis reason, we will define various treatment and comparison groups.

For defining treatment, in our preferred approach, we rely on the extrapolated data andfocus on individuals that are reasonably believed to be paying wealth taxes under the 700,000Euro filing threshold in place in most regions from 2011-2015. We classify an individual asbeing in the treatment group if their taxable wealth in 2010 is estimated to be above 700,000Euro. We refer to this group as the “2010 wealthy.” Note that the extrapolation is onlydone for individuals filing wealth taxes in 2007 because for the rest of filers, we have nobaseline wealth information. The advantage of this approach is that the treatment is basedon the immediate year prior to the reintroduction of the wealth tax, but with the limitationof using extrapolated rather than observed wealth tax data.14

A second approach defines the treatment sample on the basis of the 2007 records, whichavoids relying on extrapolated data. We classify an individual as treated by the decentral-ization if they filed wealth taxes under the centralized regime in 2007 and had a wealth ofmore than 700,000 Euro in 2007. We refer to this group as the “2007 filers.” Using theadministrative wealth tax data to determine who has more than 700,000 Euros in 2007 onlyclassifies 4% of individuals differently than using extrapolated 2010 wealth.

For the comparison group, our preferred specification includes anyone who reports positivedividends on their personal income tax form at least once when the wealth tax was suppressed,but did not file wealth taxes in 2007. In 2007, Spain introduced an exemption of up to 1,500Euro on dividends, so that this group only includes individuals that have more than 1500Euro of dividend income. We refer to this group as “High dividend non-filers.”15 Thisis our preferred group because they have a significant amount of savings, but not enough sothat they would move in response to expected wealth tax increases.

As a second approach, we use all personal income tax filers that were not wealth tax filersin 2007 as a comparison group. We name this the “2007 non-filers” comparison group.

14If wealth taxpayers illegally hide a substantial share of their taxable wealth under the centralized regime,we would not observe this wealth in tax records and could mismeasure their “true” treatment status.Nonetheless, given that there is third party information reporting on nearly 90% of total taxable wealth (i.e.commercial and residential properties, land, and financial assets deposited in domestic banks), misreportingif anything should only have a minimal effect on treatment status.

15Here being a non-filer refers to year 2007. While it is unlikely anyone in this group could become a filer insubsequent years, it is not entirely impossible, as individuals might receive a large bequest.

11


Finally, for regressions using total wealth rather than the total number of filers, we use 2007wealth tax filers that have a level of wealth that is sufficiently below the new 700,000 Eurothreshold as a third comparison group. We assign individuals to the comparison group iftheir 2010 (or 2007) wealth is between 108,182.18 and 300,000 Euro.16 Given some of theseindividuals may expect their wealth to grow and be subject to the tax, we may underestimatethe the effect on total wealth as a result. We call this comparison group the “

average baseline probability to move for wealthy individuals (vertical dashed lines), which is3.2% before and 3.5% after decentralization, and the differences in the probability to move byage, gender, income source, and position in the wealth tax schedule. We find little group leveldifferences. We find a larger probability to move among those located in higher wealth-taxbrackets, where the tax differential to Madrid is largest.


3.2 Aggregate Analysis

3.2.1 Identification Strategy

To study the effect of Madrid’s status as a tax haven on mobility, we next constructaggregated tabulations from the personal income and wealth tax micro files. We focus on thestock of wealthy taxpayers rather than wealth, as we directly observe their fiscal residenceacross time. However, we rerun the analysis using the stock of wealth as a robustness check.In our preferred specification, we aggregate the counts focusing on individuals that appearin the personal income tax data for all years from 2008 to 2015. Nonetheless, we also presenttrends for a longer balanced sample covering the period 2005-2015 to show that results arenot affected by the onset of the 2008 financial crisis.18 We total the number of individualsand the amount of wealth by region, year and treatment-comparison group by tracking whereeach individual reports her fiscal residence.

We rely on the following event-study design to carry out the aggregate empirical analysis.Let r index the region, t index time and Mr be an indicator equal to one for the regionof Madrid, which sets no wealth tax rate, and zero for all other regions. In this way, wecompare the relative evolution of the number of wealthy individuals, Nrt, in Madrid relativeto all regions other than Madrid before and after decentralization. We estimate the followingequation:

lnNrt =Mr ·[ −2∑y=−5

θy ·1(y = t− 2011)+4∑y=0

βy ·1(y = t− 2011)]+Xrtα+ ζr+ ζt+ νrt, (1)

where the indicators 1(y = t− 2011) are dummies for each event year y prior to or after thereinstatement of the wealth tax and the year prior to the reform is omitted. Then, θy gives

18As the sample of individuals is balanced, regressions using the share or the number of individuals in a regionare identical. We do not use an unbalanced sample because when taxpayers are added to the panel, theyare meant to be representative of the region-income distribution and not of the region-wealth distribution,so that we end up with a less representative sample of wealth tax filers (e.g., younger, lower wealth).

13


the evolution of the number of wealthy individuals in Madrid relative to other regions in theyears prior to 2010, while βy represents the evolution following the reform. The vector Xr,tcontains controls that includes public spending on various programs, regional demographics,amenity and economic controls, while ζr and ζt are region and year fixed effects.19

As supporting evidence of our identifying assumptions, θy should be close to zero. Inparticular, as tax policy is not set randomly across regions, any state-specific unobservablethat is correlated with taxes and the mobility of wealth tax filers may confound our results.A positive treatment effect for Madrid would indicate βy > 0 for wealth tax filers. Becausethe regressions involve the stock of individuals and not the flow, we expect βy to increasegradually rather than to jump on impact. As in Akcigit et al. (2016), we assume that thetax rate of any one other region only has a negligible impact on the number of wealthy inregion r.

The most relevant threat to identification would come from a shock that makes Madridrelatively more attractive compared to other regions. We thus add an additional layer ofdifferencing via the comparison group in each region year in a triple interaction design. Letf = T,C index the treatment and comparison groups defined in section 2.3, respectively. Wecan then define an indicator variable Wf that equals one for the treatment group and zerofor the comparison group. We estimate:

lnNrft = Wf ·Mr ·[ −2∑y=−5

θy · 1(y = t− 2011) +4∑y=0

βy · 1(y = t− 2011)]

(2)

+Xrtα + ζf + ζr + ζt + νrft,

where Xrft now includes all interactions of Wf , Mr, and year dummies and ζf are treatmentgroup fixed effects. This added difference removes any common changes that also affectthe comparison group, such as other state policies, economic conditions, or amenities thatmay have made Madrid a more attractive place for high wealth individuals. We cluster thestandard errors at the regional level to allow for an arbitrary correlation within region overtime. Given Spain has only seventeen regions (clusters), the variance matrix estimate willbe downward-biased. We follow Cameron and Miller (2015) and implement the percentile-twild cluster bootstrap, imposing the null, in order to present accurate p-values.

Given that migration to Madrid is critical, due to its zero tax status, the prior approachusing Madrid as a treatment indicator is justified. However, other tax differentials between

19These time-varying regional covariates include unemployment, GDP per capita, long term unemployment,R&D spending, poverty, high school and tertiary education, gender, median age, fraction of elderly, fertilityand mortality rate, heating and cooling degree days, and public spending on the most important governmentservices. We show results are robust to the exclusion of controls.

14


regions may matter and we can more adequately model the tax differential between Madridand other regions. To obtain an elasticity of the stock, we estimate

ln(Nrt) = � · ln(1− τrt) + ζr + ζt +Xrtα + νrt, (3)

where Nrt is the number wealth tax filers (or amount of wealth) in region r in year t, 1−τrt isthe wealth weighted net-of-average-tax rate, and all other variables remain the same. Becausethe net-of-tax rate is close to 1, the coefficient � can be interpreted as a classical elasticityor alternatively, � is (approximately) the semi-elasticity corresponding to a one percentagepoint change in the net-of-tax rate. In addition, we can augment the design to include region-time data for both the treatment and comparison group. To do so, we add all appropriateinteractions with the treatment indicator Wf and estimate the coefficient on Wf · ln(1− τrtf ).

As moving is an extensive margin response, the decision to move to Madrid is based offthe average tax rate (ATR). We first simulate the ATR for every wealth tax filer in everyregion and year, using their time-varying wealth and the tax calculator discussed in section2.2. As we use aggregate data, we then construct the mean ATR as a weighted average acrossall individuals, where following Smith et al. (2019a) we weight by the amount of 2010 wealth.

To address measurement error concerns and possible endogeneity resulting from taxablewealth changing over time, we instrument for ln(1− τrt). We do use by using the mechanicalnet of average tax rate ln(1− τrt), that is, the simulated rate holding wealth constant at its2010 level. This latter tax rate uses only statutory variation in the ATR. As an alternative,we also follow Kleven and Schultz (2014) and instrument with the binary Madrid × Postvariable. The use of these two instruments provides local average treatment effects (LATE)for two different sub-populations that provides us with some intuition of which regions drivethe effects. In the case of Madrid × Post, the instrument only induces a change in the taxof Madrid relative to other regions. In this way, we think of the LATE interpretation asidentifying the effect of Madrid’s non-adoption of a wealth tax. When we use the simulated1− τrt instrument, matters are more complex because the instrument is continuous. In thiscase, a change in the instrument induces a change in the tax rates of all regions. Thus, theelasticity using this instrument is with respect to all tax differentials within Spain.

3.2.2 Results

Figure 5 shows θy and βy from estimation of (1). We present separately estimated coef-ficients for the treatment and comparison groups, so that the reader can observe the trendsin both the treatment and control group. The upper panel uses our preferred comparisongroup, “High dividend,” while the lower panel uses the “2007 non-filers.” The left and right

15


panels present results using the balanced 2005-2015 and 2008-2015 samples, respectively.20

for the “2010 wealthy,” the number of filers located in Madrid steadily increases followingdecentralization. The relative stock of wealthy individuals becomes statistically differentthree years after decentralization and by five years after the reform, Madrid’s relative stockof wealthy individuals increases by 11%. Although the relative stock of wealthy individuals inMadrid increases in the two years after the reform, these result are not statistically significantfor two main reasons. First, although migration flows may jump on impact, the stock is aslower moving variable. Second, the first two years of decentralization were characterizedby a large amount of uncertainty and a retroactive application of the tax, which may havehindered any type of tax reoptimization via a change of residence. In subsequent analysis,we focus on the shorter balanced sample, which as we have already mentioned, is morerepresentative of the wealthy population.21

In support of our identifying assumptions, find little pretrends in the relative attractive-ness of Madrid to other regions. Critically, θy being close to zero shows that mobility effectsfollow tax changes and do not predate them such that there are no pretrends in the periodsprior to the reform. Moreover, the common pretrend of the comparison group allows us toadd another layer of differencing.


Although the comparison group shows a minor upward trend following the reform, thisincrease is statistically insignificant and will only result in slightly smaller estimates using(2). Moreover, this suggests that it is unlikely there are unobservable factors making Madrida relatively more attractive region to wealth tax filers. In Table 1 (Panel I), we present asimple design that usesWf ×Mr × Post rather than the generalized (dynamic) design above.This simpler specification identifies an average effect across all post-reform periods, whichgiven the dynamic effects noted above, will understate the cumulative effect. For this reason,in Panel II, we also present the cumulative effect given by the coefficient on the interactionwith the Madrid dummy and the year dummy for 2015 from the estimation of (2). Consistentwith the event study figures above, estimating (2) using the “High dividend” or the “2007non-filers” comparison groups only lowers the effects by a small amount.

20Figure A4, Table A3, and Table A4 show the results are robust to the use of the “2007 filers” treatmentgroup. As expected, because this latter group is only different for individuals near the filing threshold, theresults are almost identical.

21Figure A5 shows the same results apply when using instead the (log) amount of wealth as the dependentvariable. Unlike the number of individuals, which is directly observed in the administrative data, these re-sults make use of extrapolated data. Results are robust to using (observed) 2007 rather than (extrapolated)2010 wealth, as seen in Figure A5.

16


[Table 1 about here.]

Table 2 presents the elasticity estimates for the number of filers and the amount ofwealth.22 Models (a/e) are estimated using OLS, while models (b/f) and (c/g) presentIV estimates using the simulated net-of-tax rate and the Madrid × Post interaction, respec-tively. With respect to the number of filers, the first instrument yields an elasticity of 5.3. Inother words, a one percent increase in the net-of-tax rate, which corresponds to an (approx-imately) 1 percentage point decline in the average tax rate, increases the number of filersin the region by 5.3%. Critically, when using the Madrid × Post instrument, the elasticityincreases to 7.9. Given the wealth weighted average tax rate across regions is approximately0.97 percent, these results are similar to our prior estimates from the event study design.Consistent with the LATE intuition above, this specification identifies tax-induced mobilityusing only the relative differential with Madrid and not the much smaller ATR differencesbetween other regions. The increase in the coefficient from the binary instrument suggestsMadrid is critical. When dropping Madrid (d/h) and exploiting only smaller tax differentials,the elasticity decreases substantially and is insignificant. Overall, we conclude that Madrid’szero tax rate plays a special role and other tax differentials do not matter as much.


3.2.3 Comparison to Income Tax Elasticities

How do these results compare to the literature on income-tax-induced migration? Wealthtaxes are a tax on the stock of wealth, while capital income taxes are a tax on the flow. Giventhe still scarce literature on the wealth tax, we convert our estimates to an equivalent capitalincome tax to allow for comparison with the larger literature on income tax elasticities.Following Kopczuk (2019), suppose that an individual with wealth W and a rate of returnR in a given year can either be taxed next year on the accumulated stock (1 + R)W or onthe return, RW . Then, a wealth tax rate τ will raise an equivalent amount of revenues as acapital income tax rate of T where the relationship is given by

T =(1 +R)τ

R. (4)

We can then convert our wealth tax elasticity, �1−τ , with respect to the wealth weighed

22Figure A6 depicts a binned scatter plot for the number of individuals for our preferred treatment andcomparison groups using the wealth weighted average tax rate.

17


net-of-tax rate, 1− τ ≈ 0.97, using

�1−T = �1−τdln(1− τ)dln(1− T )

, (5)

where �1−T is the elasticity with respect to the net-of-tax rate on capital income given by (4).Figure 6 indicates that the magnitude is remarkably similar to the literature on the

mobility of top income earners. Using the average rate of return for the top 1% wealth group,we estimate an income tax elasticity of approximately 0.34. When excluding Madrid to relyon only smaller tax differentials between states, this elasticity falls to 0.1. Critically, ourestimates represent short-term to medium-term responses, and when available, we comparethis to estimates for the same time horizons of income tax studies in the figure.23


3.3 Individual Choice Model

3.3.1 Identification Strategy

We complement the aggregate results with an analysis at the individual level by meansof a location choice model. This allows us to control for individual-specific factors that mayinfluence the probability of moving to – or residing in – a specific region, to account for regionby year fixed effects, and analyze if the effects are heterogeneous across groups of individuals.We use again the “2010 wealthy” sample as our preferred treatment group, but we show thatthe results are robust to the use of the “2007 filers” sample. Unlike the aggregate analysis,we do not need to balance our sample, which allows us to test if results are sensitive to doingso.

For our purpose, a “move” or “stay” (we refer to these as a case) is an individual time-specific event that is indexed by i and t and the regions in the choice set are indexed byj. We will focus on two samples: the full and the movers sample. The full sample is thesame we use in the aggregate analysis and includes both, movers and stayers. The moverssample includes all individuals that relocated across regions between period t and t− 1.24 If

23One exception is Young et al. (2016) who estimate a long-term response. The elasticity reported for Morettiand Wilson (2019) is a short-run elasticity; however, these authors also estimate the effect of a permanentone percent increase in the net-of-tax rate between year t and t + 5 would lead to a 6.0 percent increasein the stock of scientists by the end of year t+ 10. Under strong assumptions, Kleven et al. (2013) reportlong-run elasticities that are only slightly larger than those in the figure. Akcigit et al. (2016) show thatdomestic [foreign] inventors long-term mobility is slightly less [more] sensitive to tax rates.

24Given movers are only a fraction of the stock, focusing on the movers sub-sample reduces endogeneityconcerns if governments were to set tax rates based on the stock of wealthy rather than on the number ofmovers (Schmidheiny, 2006; Brülhart et al., 2015).

18


an individual moves more than once, each move represents a case.The dependent variable ditj is equal to one for the chosen region of fiscal residence in year

t and zero for all other regions. In other words, it equals one for the destination region ifthe person moved or for the region of residence if the person stays. In its simplest form, weestimate the following linear probability model:

ditj = βMj × Postt + ζjzit +Xtjα + ιj + ωit + εitj, (6)

where Mj is an indicator equal to one for Madrid and zero for the other regions; Postt isanother indicator equal to one for the years following decentralization and zero otherwise; βcaptures the effect of the zero-tax regime of Madrid from 2011 onward; ζjzit are interactionsof region dummies with characteristics of the taxpayer (i.e., gender, age, age squared, genderby age, and labor income),25 which makes it possible to estimate a region-specific return foreach covariate and to flexibly allow for wealth accumulation to differ across regions betweenmen and women and by age; Xtj are the same controls used in the aggregate analysis at theregion-year level, and ωit are fixed effects at the case level. The inclusion of ωit is crucialbecause it forces identification of our parameter of interest based on within-case variationacross alternative regions for a specific taxpayer in a given year. Thus, identification followsfrom the fact that, for each case, Madrid has a tax differential with all other choice regions.The specification allows for alternative fixed effects ιj, that control for all time-constantcharacteristics of a specific region. The Madrid dummy Mj is one of these fixed effects.

We use a linear probability model.26 This is based on our desire to include many binarycovariates for which logit models are ill-suited, along with our desire in future specificationsto instrument for the tax rate. Although the probability of any one region is not bounded inthe linear model, the ωit forces the predicted probabilities over all regions to sum up to onefor each individual in a given year.27 For this reason, an increase in the predicted probabilityof one region must necessarily decrease the probability of choosing other regions.

The specification of (6) results in a difference-in-differences interpretation embedded ina location choice model, which can be generalized by interacting the set of alternative-fixed

25Note that all variables are individual specific, but do not vary across alternatives. The interaction ofcharacteristics with region dummies allows for a different coefficient for each potential region of choice.

26The specification of (6) is the linear equivalent to an alternative-specific conditional logit.27The fact that the linear probability is not bounded between 0 and 1 is not a problem given we care aboutthe partial effect of taxes on the dependent variable, and not the fitted probability per se. The advantage ofa nonlinear framework is the ability to relax the IIA assumption. Given most mobility is driven by Madrid,the odds of choosing Madrid over Catalonia, for example, are unlikely to differ when the alternatives includeor exclude different regions. In a theoretical model below, we show this is true or any bias is likely minimal,so that the linear probability approach is suitable and comes with many advantages for our setting.

19


effects ιj with year dummies.This specification allows us to integrate an event-study approachinto the location-choice environment by replacing the Post dummy with year-indicators inorder to to estimate dynamic treatment effects. We estimate:

ditj =

[ −2∑y=−5

θyj · ιj 6=ĵ·1(y = t− 2011) +4∑y=0

βyj · ιj 6=ĵ·1(y = t− 2011)]

(7)

+ ιj + ωit + ζjzit +Xtjα + εitj,

where y stands for “event time”. Such a specification requires omitting a given region ĵ and agiven year, which is the year immediately prior to the reform. This more general specificationreveals the location choice pattern for each of the regions relative to the omitted region andyear immediately prior to the reform. Coefficients βyj for j = Madrid capture the differencein the probability of choosing Madrid after the reform relative to a baseline region and theyear prior the reform. The coefficients θyj for j = Madrid show that same relative evolution,but for a period where taxes do not differ. The specification is more flexible than a simpleinteraction with only a Madrid indicator because it allows us to estimate the effect of smallertax differentials between any of the other regions.

We complement these results with estimations using the net-of-average-tax rate, whichwe simulate using person-specific wealth in every year t for each taxpayer and all alternativeregions j, as described in Section 2.2. We then estimate the location choice model using thenet-of-average-tax rate as the main independent variable:

ditj = β · ln(1− τitj) + ωit + ρtj + ζjzit +Xtjα + εitj. (8)

The coefficient on the (log) net-of-tax rate represents the change in the probability of movingto (or staying in) a region for a one-percent – approximately a one percentage point – changein the net-of-average-tax rate. We complement the OLS regressions with IV estimates wherewe follow the approach of the aggregate analysis and use the net-of-tax rate based on anindividual’s 2010 pre-reform tax base as an instrument.

The specification based on tax rates comes with an added advantage: because the taxsystem is progressive, we have variation of tax rates across individuals within a region-year.Thus, we can include region by year fixed effects, which account for other contemporaneouspolicy choices that a region may make. These region-year fixed effects, ρtj, also accountfor any unobserved time-varying economic shocks or amenities that influence the relativeattractiveness of a given region. However, inclusion of these ρtj comes with a cost. IfMadrid’s status as a tax haven plays a special role, then some of this effect will be absorbedin the region-year fixed effects and may result in an underestimation of the true effect. For

20


this reason, we also present results excluding region-year fixed effects.We cluster standard errors at the origin-tax-bracket and alternative-tax-bracket level,

following previous studies (Akcigit et al., 2016; Moretti and Wilson, 2019), which cluster atthe origin/destination-ability level. In our setting, the wealth tax brackets form analogouspartitions to the ability partitions. Using the region of origin to calculate standard errorsimplies that we cannot use the first year of the sample, 2005. Thus, the sample used here isthe same as in the aggregate analysis, but the event studies start in 2006.

3.3.2 Baseline Results

Table 3 shows the results from the estimation of (6), adding controls sequentially. Panel Ipresents estimates of the model for the full sample between 2006 and 2015, for all individualsin the “2010 wealthy” treatment group, including movers and stayers. The magnitude of theeffect of Madrid is 0.016 in our preferred specification with alternative-region controls andalternative fixed effects, as well as individual controls, but results vary little across specifi-cations. Given the baseline probability of residing in Madrid in the pre-reform period was22.3%, our model suggests that following decentralization, this is equivalent to an approxi-mately 7.2% increase in the share of wealthy individuals in Madrid. This effect is statisticallysimilar to Panel I of Table 1 estimating the average post-reform effect. For a wealth weightedaverage tax rate, the resulting elasticity would be similar to previously.


Panel II of Table 3 presents the results from the same estimation as in Panel I, but forthe sample of movers. As expected, the magnitude of the coefficient conditional on movingincreases substantially compared to the full sample. Across all specifications, the resultsindicate that the probability of changing the fiscal residence to Madrid from any other regionin Spain increased after decentralization by about 23 percentage points. To benchmark thisnumber, conditional on moving, the baseline probability of moving to Madrid in the pre-reform period was 46.2%. This estimate represents the effect of the mean tax rate differentialbetween Madrid and other locations, which for movers, is approximately 0.44%.

The use of individual data allows us to perform robustness checks that rule out potentialconfounders, as discussed previously. In particular, results are robust to alternative treatmentsamples and to dropping individuals who may be influenced by labor income taxes.28

28Table A5 in the appendix estimates the model using the “2007 wealthy” treatment sample and excludingindividuals with labor income above 100,000 Euro – the threshold above which income tax differentialsbecome important after 2010. Results in both cases are almost identical to those in Table 3 for the full andthe movers sample, suggesting that the effect is driven by changes in the regional wealth tax design andnot my measurement error due to extrapolation or changes in personal income taxes.

21


Figure 7 shows the annual estimates from (7) for the full sample and the movers sample,but interacting the event year dummies with the Madrid dummy, Mj, instead of with regionspecific indicators, ιj 6=ĵ. The interpretation of the coefficients is thus the difference over time(relative to 2010) of choosing Madrid relative to all other regions. We find no major differencesprior to the reform, which confirms that the effect is indeed driven by the decentralization ofthe tax.29 Moreover, the event study demonstrates a clear trend break following the reform,which would persist if adjusted for the minor trends in the pre-period. The cumulative effectof the reform is obtained by looking at the final event dummy coefficient of the full sample,which represents a 0.024 percentage point change in the probability of choosing Madrid asthe fiscal residence. Given the baseline probability of selecting Madrid is 23.7% in 2010, theprobability rises to 26.1% five years after decentralization. This represents an 10% change inthe stock of filers, comparable to the prior aggregate analysis.

The unbalanced and balanced sample are nearly identical, which suggests that non-random attrition, perhaps due to death, non-filing, or out-of-country migration, does notthreaten our results. All further results are based on the unbalanced sample, but are robustto using the balanced data. Even though our data do not allow for entry of new wealth taxfilers from abroad, we find based on aggregated data provided to us by the tax authority,that the number of wealth tax filers that entered Spain from abroad is small relative to thewithin country flows.


As an alternative source of identification, we also exploit the individual-specific tax ratesfrom our calculator by estimating (8). Table 4 presents the results from this estimation forthe full sample using OLS and IV models, using the simulated mechanical tax rate, 1− atritjas an instrument (Panel I and II, respectively), and the same estimations for the sample ofmovers (Panels III and IV).

Column (a) includes alternative fixed effects only, while column (b) includes the samecontrols as in the full specification of Table 3. Given variation of average tax rates withinregions across the wealth distribution, we can additionally include a dummy variable for eachalternative j in each year t. Column (c) presents results with alternative region-year fixedeffects, forcing thus identification from the variation of relative differences of average taxrates within region-year pairs. This specification is useful to address concerns about time

29One might be concerned that the results are sensitive to the use of the Madrid dummy relative to all otherregions. In an exercise similar to a jackknife procedure, Figure A7 in the appendix estimates (7) sixteentimes for each possibly omitted region. It shows that results are not sensitive to the omitted region or togrouping all regions into a single counterfactual category.

22


varying region specific shocks or changes in amenities, as well as any other fiscal instrumentwhich might change in a single region and affects all taxpayers in that same region. Column(d) adds individual controls. Column (e) drops individuals selecting Madrid.


For the full sample and using OLS, a one-percent increase in the net-of-average-tax-rateincreases the probability of residing Madrid by 8.2 percentage points. Both OLS and IVestimates are similar, suggesting that most identifying variation comes from statutory taxrate variation.30 For movers, a one-percent increase in the net-of-average-tax-rate increasesthe probability of declaring Madrid by 6.5 percentage points. While point estimates are notsubstantially different to the full sample, the average net-of-average tax rate in the full sampleis only 0.24% and almost twice as large for the sample of movers. This relates to our previousfindings that individuals in higher tax brackets have a higher probability of being a mover.Furthermore, the probability of choosing Madrid in the full sample (23.3%) is different to theone in the movers sample (59.7%).

To show more formally the special role of Madrid, column (e) drops movers to and stayersin Madrid from the analysis such that the effects are only identified based on the smaller taxdifferentials between regions other than Madrid. The estimated coefficient is approximatelyone-tenth the size of the prior results for the full sample and falls even more for the movers.All estimates turn insignificant, suggesting that the differential of any region relative toMadrid’s zero tax rate is critical for the mobility effect.

3.3.3 Heterogeneity

To analyze heterogeneous responses, we exploit characteristics of individuals, such as ageand gender, and construct categories related to their their financial situation before 2011.We differentiate between individuals that filed non-incorporated business income, dividendincome, effective rents from tenant occupied housing, and imputed rents from owner-occupiedhousing in any year from 2008-2010 in her income tax declaration. Interacting indicators forthese variables with our Madrid × Post-term allows us to see if Madrid’s status as a taxhaven prompts heterogeneous responses by individual characteristics.


30Table A6 shows results are robust to the use of the simulated tax rate based on the “2007 wealthy” treatmentsample and the Madrid × post interaction as an instrument. IV estimates using the binary instrumentincrease for the sample of movers consistent with the complier intuition discussed in the aggregate analysis.

23


Results regarding personal characteristics are shown in Figure 8(a) for the full sampleand 8(b) for the movers sample. We do not find substantial heterogeneity in the effects.Results suggest that age and gender do not matter for the magnitude of coefficients.31 Withrespect to life-cycle effects, the lack of timing moves before/after retirement suggests a lackof forward-looking (forecasting) behavior by households. Furthermore, we provide estimatesbased on the composition of asset portfolios. Again, no significant difference emerges betweendividend and business owners. Movers with real estate are slightly more responsive, however,most individuals in the sample own some real estate. The result for owner occupied realestate is driven by the fact that 88% of our sample declare income from that asset type.

To analyze how effects vary over the distribution of wealth, panels 8(c) and 8(d) showresults by the individual’s top bracket of the tax schedule. For the full sample, panel 8(c)shows substantial variation with the largest effects in higher brackets. Panel 8(d) showslittle variation across movers.32 This confirms that the overall effect is driven by morewealthy individuals moving to and staying in Madrid. Apart from these effects, the lack ofheterogeneity in the sample of movers is explained by the fact that high wealth individuals arerelatively homogeneous in their tax avoidance preferences. Instead, the heterogeneity arisesover whether to move or not (Figure 4) rather than the probability of choosing Madrid.

3.3.4 Additional Evidence on the Special Role of Madrid

Tables 2 and 4 already provide some initial evidence that most of the mobility is due to thetax differential with one region: Madrid. However, there also exist smaller tax differentialsbetween other regions that may potentially lead to wealth tax filers changing their fiscalresidence from one region to another. To trace out pairwise effects, we estimate (7) seventeentimes, omitting a different region each time. This flexible specification allows us to plotsimilar graphs of the mobility responses for all region pairs to a baseline region differentthan Madrid. As an example, Figure A8 shows the mobility effect of all regions relative toCastile-La Mancha. Only the region of Madrid shows a significant pattern, while all otherregions so no pairwise effect. This confirms that all mobility responses are indeed driven bymoves between Madrid and other regions, but not between them.

We repeat this exercise for every region in Spain to show that the null results for non-Madrid regions generalize to every possible omitted region, and thus all region pairs in Spain.

31The fact that the effect does not increase in age (the point estimate for individuals above 80 is even lowercompared to younger individuals) reassures us that moves are not motivated by other tax instruments, suchas inheritance taxes, although as noted previously the inheritance tax provides no additional incentive tomove starting in 2011. Only 9% of movers are 80 or older in this sample.

32The small effect of the top bracket might be driven by the fact that only 2.81% of movers are in this bracket.

24


Figure 9(a) shows the aggregated post-reform effect for the sample of movers (the populationrelevant for the theory discussed subsequently) that results are driven by pairs involvingMadrid. All regions see a decline in the probability of moving there relative to Madrid (reddiamonds). Only pairs involving Madrid as a destination see an increase in the probability ofmoving (blue circles). Almost all other pairs not involving Madrid show insignificant effects.


Although informative, a concern is that even if all other region pairs have small effects,the difference in taxes between Madrid and the other places is so large that the effect scaledby the tax change is actually homogeneous. To address this, in Figure 9(b), we re-estimate(8) excluding movers to single destination region at a time. Critically, when we excludemovers to Madrid (red diamond) as we did in Table 2 and 4 previously, the effect of thetax differential between regions becomes zero. However, this is not the case when we dropmovers to any region other than Madrid, and none of those estimates is statistically differentfrom the baseline estimate as indicated by the red dashed line.

Overall, these exercises reveal that the zero tax in Madrid is critical. In other words,inter-jurisdictional wealth tax differentials, when small, appear not to matter in the locationchoice decisions. However, the fiscal residency is intensely affected by the presence of atax haven that facilitates dramatic tax evasion. These results are critical to the subsequenttheoretical model we will develop.

4 Evasion vs. Migration: Theory and More Evidence

The moves we see in the data may be tax avoidance (real migration that reduces taxliability) or tax evasion (fraudulently declaring fiscal residence).33 Although it is not possibleto causally disentangle whether these responses are real or not, the fact that Madrid is drivingthe results is quite revealing in terms of the underlying model of mobility that can be usedto rationalize our findings. The basic intuition is that in a standard mobility model, evena small decrease from a positive tax rate will attract some marginal individuals. That isnot the case in our results: taxpayers appear to be aiming for the lowest possible tax rate.Hence, this evidence alone suggests that our findings reflect reporting/shifting responses and

33An example of real migration could be if a wealth taxpayer living in any other region but Madrid wouldbuy an apartment in the capital (or already have a second home) and move there to avoid the wealth tax.An example of evasion could be if instead this same person would buy the apartment (or would alreadyhave a secondary residence) and pretend that she lives there but stay in the same region. This latter effectcould occur by simply falsely misreporting the number of days spent in the primary/secondary residenceor even declaring their primary residence as the address of a relative.

25


not real migration. We formalize this in a simple model and provide additional empiricalevidence using regional variation in audit rates supporting the evasion channel.

4.1 A Model of Migration and Evasion

An individual i endowed with wealth W i lives for two periods: prior to decentralization(t = 1) and after decentralization (t = 2). Prior to wealth tax decentralization, the indi-vidual chooses to reside in, without loss of generality, region h. After decentralization, theindividuals makes a new choice. Let j index the regions of Spain: j = h is the home region,j = m is Madrid, and j = 1, ..., J are the alternatives other than Madrid. Taxpayers make adecision on where to live and which region to declare as the fiscal residence.

Consistent with the data, we assume this high-wealth individual is a rentier and consumesonly her capital income. Given a global market for capital and thus a world rate of return,Rt, this implies pre-tax consumption cit = RtW i. As noted previously, an annual wealth taxτ ijt is equivalent to a capital tax T ijt given by (4). Thus, we use this tax to solve the model.Absent moving costs, the utility from individual i choosing region j in time t is given byu(cit(1− T ijt), zijt) = cit(1− T ijt) + g(zijt), where z are amenities in the region of residence.34

Starting from the standard model of tax evasion (Allingham and Sandmo, 1972) and atraditional model of migration (Akcigit et al., 2016), we make two theoretical contributions.First, we modify the standard tax evasion model, which traditionally involves the taxpayerselecting the amount of income to hide from the authority, to allow for the taxpayer to makea discrete all-or-nothing decision. In making this decision, the taxpayer must choose amongmultiple taxing jurisdictions when deciding where to shelter her wealth. Second, we combinethe standard mobility and evasion models, such that the taxpayer has the choice over evadingversus migrating, along with which region to evade or migrate. In other words, the taxpayercan shelter (via evasion) all of her income in a lower-tax region at some expected cost butmaintain the amenities of her home region or can migrate (via a real move) to the regionat some cost that also results in giving up the home region amenities. In order to buildintuition, we first consider the cases with only real migration or only evasion.

34We assume a quasi-linear utility function, which implies that the taxpayer is risk neutral and moving costsdo not incur income effects. As will become clear, a small perturbation making the taxpayer risk aversewill not change results. Evasion results will hold if the coefficient of risk aversion is sufficiently small.

26


4.1.1 Migration Only

Let us first consider the standard model of migration where an individual can only move.In this case, region j will be chosen after decentralization from the set j′ = {m,h, 1, ..J} if

u(cit(1− T ijt)− φihjtcit, zijt) = argmaxj′

{u(cit(1− T ij′t)− φihjtcit, zij′t)

}, (9)

where moving costs are given by φihjtcit with φhjt < 1 and φihht = 0.35 The model makes itclear that the probability that an individual located in a given region depends on the fullvector of taxes in all regions. Thus, a marginal decrease in the tax rate of any one region,for example region J , relative to the home region will induce added migration to that regionfor individuals if u(cit(1 − T iht), ziht) − u(cit(1 − T iJt) − φihJtcit, ziJt) was small prior to the taxdecrease and region J was the next best alternative. Hence, because the migration decisiondepends not only on the tax differential but also on the amenities and the moving costs tothe destination region, the model predicts that not all migration is to Madrid.

4.1.2 Evasion Only

Next, we modify the Allingham and Sandmo (1972) and Yitzhaki (1974) model of taxevasion such that the taxpayer makes-an-all or nothing decision to shelter their wealth andmust select which region to shelter it. In our model, an individual chooses a region j todeclare taxes, so Tjt depends on the region of choice. However, with evasion, the individualcan stay living in the home region h and so local amenities are given by the home region,ziht. Moreover, tax evasion is risky and the individual faces a probability of being caughtof pi ∈ [0, 1] and a fine f i.36 As in Dharmapala (2016), the individual incurs idiosyncraticcosts, κitcit, κit < 1, of evasion because individuals have internalized norms of tax complianceto varying degrees. Then, the utility of declaring one’s home region as the fiscal residence iscit(1 − T iht) + g(ziht) and the utility declaring any other region j 6= h is (1 − pi)cit(1 − T ijt) +pi[cit(1− T iht)− f i(T iht − T ijt)cit

]− κitcit + g(zihit), where if an individual is caught, they must

pay all taxes due and a fine that is proportional to the amount of income evaded. With allderivations in Appendix A.3, evading in Madrid is preferred to truthfully reporting the homeregion if

T iht(1− pi − pif i)1− pi

> κit. (10)

35The moving cost (and the idiosyncratic evasion cost introduced later) are modeled as a share of the pre-taxcapital income flow. Given they are person-specific, they can also be written in dollars, but the percentformulation facilitates comparison to standard tax evasion models.

36By 305 Codigo Penal and 192 Ley General Tributaria (LGT), the fine is a percent of taxes hidden (Yitzhaki,1974). The Allingham and Sandmo (1972) penalty function would lead to even starker results.

27


Note that if the idiosyncratic costs are zero, as in the standard evasion model, this expressionis always true if pi < 1/(1 + f i) and implies Madrid is preferable if the audit probability issufficiently small. Under Spanish law, the fine is approximately 100% for most individuals inour sample, but higher at the top, which implies pi < 0.50.

Then, if pi is sufficiently small, unlike the migration model, then Madrid will always bechosen for tax evasion. The intuition can easily be seen in the limiting case where pi → 0. Asthe audit probability approaches zero, the form of the fine is irrelevant, and the individualwill simply evade in the region that affords them the largest benefit from tax savings.

4.1.3 Evasion and Migration

Finally, we consider the most realistic scenario in which the taxpayer has choice overmigrating or evading.37 Focusing on evasion, evading in Madrid will be preferable to movingto Madrid if

− T ihtpicit − pif iT ihtcit > g(zimt)− g(ziht) + κitcit − φihmtcit. (11)

If pi → 0, only the differences in the valuation of amenities and evasion/moving costs matter.By revealed preference, in the pre-decentralization period, the home region was chosen

over Madrid, which means that g(zimt) − g(ziht) − φihmtcit < 0 for t = 1. Consider the casewhere κit = 0. If amenities and moving costs are time invariant (approximately similar) inboth periods, the right side of (11) is negative and evading via Madrid is always optimal aspi → 0. Moreover, if the valuation of amenities in both regions is the same, but changingover time, this term is also negative and evading is the better option if κit is sufficiently small.More generally, the sufficient condition for evasion via Madrid to dominate moving is thatthe audit probability and idiosyncratic evasion costs are sufficiently small. If this conditiondoes not hold, no evasion will occur and individuals may move to Madrid or any other region.

Proposition 1. If the probability of detection and idiosyncratic evasion costs are sufficientlylow, all fraudulent changes of fiscal residence will be to the tax haven and any increase in thestock of taxpayers in non-havens must be due to real moves.

The proposition sheds light on our empirical results. Given in Figures 9(a) and 9(b) wefind the stock of taxpayers only increases in Madrid and not in other regions with (positive)low tax rates, taxpayer migration is likely limited. Our theoretical model suggests such a

37It is also possible that an individual moves from their home region to a region other than Madrid, butsimultaneously falsely declares Madrid. If the person simultaneously evades, then taxes between the homeregion and new residential region are irrelevant for the real move and so a real move would only arise ifamenities change dramatically over time. Such dramatic change is unlikely, and even if it did arise, it wouldsimply mean a minor modification to the necessary audit probability threshold.

28


corner solution is consistent with a reporting/shifting response than a real relocation. Asaudit probabilities and costs of evasion are person-specific, both tax evasion and real movesmay exist simultaneously. Nonetheless, given the very small audit probabilities we find inthe next section, evasion is likely the dominant mechanism.

4.2 Audit Rates and the Evasion Channel

Standard tax evasion models assume that the aggregate audit rate, p, increase with eva-sion, e, so that p′(e) > 0 (Slemrod, 2019). In our context, if evasion is mainly due to movesto Madrid and not to other regions, we should expect audit rates to increase with the numberof movers to Madrid, m, but not with the number of movers to other regions, n, so that theanalog assumption the the standard model is p′(m) > 0 and p′(n) = 0.38

To test this commonly believed assumption and shed further light on the mechanisms ofmobility, we digitize tabulations on wealth audit records for each region in Spain from 2005-2015 published by the General Inspection Department of the Spanish Ministry of Finance. Anaudit can be conducted due to the misreporting of fiscal residence or any other misreportingactivity. These statistics are an upper-bound of the audit rate for fiscal residence.

Figure 10(a) shows the average annual audit rates by region before and after the decen-tralization of the wealth tax. We define the audit rate as the number audited returns dividedby the total number of wealth tax returns filed. Prior to decentralization, despite the regionsadministering and receiving wealth tax revenue, there was little regional variation in auditrates and they were less than 0.1% for nearly all regions. However, after decentralizationaudit rates increased in most regions but not in an uniform manner, ranging from 0.01% inAragon to 1.5% in Castile-La Mancha.


We analyze whether the non-uniform change in audit rates are related to evasion viadeclaration of a fraudulent residences by regressing the pre/post-reform change in auditrates on the change in the share of movers to Madrid and, separately, the change in the sh

Date post:	13-Feb-2021
Category:	Documents
Upload:	others
View:	1 times
Download:	0 times

Paraísos Fiscales, Wealth Taxation, and Mobility...2021/01/20 · acknowledges ﬁnancial support...

Documents